Perplexity's Brain Wrote 80,000 Agent Skills for 20,000 Users
Perplexity's Brain wrote over 80,000 personal skills for 20,000+ Computer agent users, with 72.9% adoption in a randomized test and an 11% drop in tool-call failures.

Why it matters
- Brain wrote more than 80,000 personal skills for over 20,000 users between August 1 and November 14, 2026.
- 72.9% of users with automatic improvement enabled received a new personal skill, versus 6.5% of controls.
- Tool-call failures fell 11% across all Computer activity; dissatisfaction threads fell 7.6%.
- Task success on the SkillLearnBench subset rose from 43.8% to 62.5% with generated skills.
- Built-in skill updates require paired replay testing and engineer approval before release.
Perplexity said on November 18, 2026, that Brain, the personalization system for its Computer agent, had written more than 80,000 personal skills for over 20,000 users between August 1 and November 14, 2026. The system now creates and updates reusable agent skills on its own, and the production data behind that capability shows why the company pushed it into default behavior: in an eight-week randomized test, 72.9% of users with automatic improvement enabled received at least one new personal skill, versus 6.5% of controls who manually added one.
The numbers matter beyond Perplexity. Agents that cannot retain workflow-specific procedures force users to repeat instructions on every session, a friction point that has slowed enterprise adoption of autonomous tools. Brain's approach packages reusable instructions, supporting resources and, in some cases, scripts into skills that Computer's harness (the runtime layer that executes the agent's actions) makes discoverable and loads on demand. Users can also invoke a skill explicitly in a prompt. The system changes the agent's context and instructions, not the underlying model's weights.
What did the production test measure?
Perplexity ran a randomized eight-week experiment across its live product. The results, attributed to the company's November 18 announcement:
- 72.9% of users with automatic improvement enabled received at least one new personal skill, against 6.5% of controls who manually added one.
- A personal skill was loaded in a user-started conversation for 53% of the enabled group and 13% of controls.
- Among customers who received a Brain-written skill, 63% used one again in a later conversation within four weeks.
- Tool-call failures fell 11% across all Computer activity, and about 7% after background runs were excluded.
- Medium- or high-dissatisfaction threads per user fell 7.6%, with no measurable change in cost or response time.
The company also reported more than 100,000 automatic updates to over 30,000 existing personal skills between September 4 and November 14, 2026. Personal skill improvement is available to all US Max subscribers, enabled by default for eligible users, and can be turned off. The Plugins tab identifies Brain-authored skills and lets users choose which ones stay active.
What does SkillLearnBench show?
Perplexity ran a narrower offline experiment on SkillLearnBench. Computer generated one skill for each of nine task families, and task success across 16 paired instances rose from 43.8% without skills to 62.5% with generated skills. The full benchmark contains 20 tasks across six real-world categories and 100 verified instances, so the company's run covered only a subset of the suite. The gain is notable. The coverage is partial, and Perplexity says so.
How does the improvement engine work?
The shared improvement engine runs four stages: gather evidence, reflect on it, verify a proposal, activate the change. The harness logs skill loads, and the agent files structured diagnostics when it detects user frustration, broken behavior or clear inefficiency. Stored session trajectories preserve messages, tool calls, reasoning and results, which lets the improvement agent inspect where a procedure went wrong rather than rely on an outcome score or a short summary.
In daily personal-skill runs, a background improvement agent receives recent sessions, skill-usage data and the existing skill catalog. It has no internet or connector access, and it can retrieve more of the user's session history as needed. The creation bar is deliberate: a new skill requires two independent occurrences of the same reusable behavior, while one grounded occurrence can support an update to an existing skill. Each proposal is limited to one change per skill per run. Deterministic checks test whether a proposal is well-formed and applicable, and an LLM judge audits sampled proposals for supporting evidence and useful instructions.
Activation writes the verified change to the user's library for future sessions. Context is a finite resource. Computer loads an index of available skills at the start of each session, so an oversized library consumes context and makes relevant skills harder to find. The system therefore caps automatically created skills at 10 per user. At capacity, the agent must retire an existing skill and justify why its replacement is more useful before adding another.
Why do built-in skills face stricter rules?
Built-in skills ship to every Perplexity user, so their updates follow a stricter path than personal ones. The process uses sanitized evidence from sessions eligible for training, stripping credentials, other secrets, attached files and media. A weekly revision job looks for failures recurring across multiple sessions rather than letting one bad session drive an edit.
A scenario-generation agent then converts those failures into self-contained tests, synthetically rebuilding unnecessary context and replacing connector-dependent inputs with synthetic files. The improvement agent receives root-cause descriptions and sanitized successful sessions, but not the held-out failure scenarios reserved for evaluation. In paired replay, every test scenario must pass with the revised skill and at least one must fail with the original. Control scenarios check that previously working behavior has not regressed.
An LLM judge scores responses and trajectories against each scenario's rubric. Candidates and their evaluation results are then submitted as an internal pull request, and a change takes effect only after an engineer approves it. Human approval remains the final gate for anything shipped to all users.
What comes next?
Perplexity plans to extend improvement to user-authored skills, skills shared within projects, and organization skills. That raises the open product question the company itself frames: how it will apply the current mix of automatic activation, replay testing and human approval as procedures move from an individual's history into shared team workflows. A skill tuned to one user's habits can fail quietly when a whole team inherits it, and the 7.6% drop in dissatisfaction threads was measured on personal use, not shared contexts. The roadmap points at organizations; the guardrails, so far, were built for individuals.
Source: unite.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
745 articles
Related articles
- Perplexity launches Teammates: AI agents that own ongoing work, not just chats
- Perplexity brings its full Computer agent to local Apple Silicon Macs
- Perplexity Launches Multiagent Mode to Coordinate Frontier AI Models
- Cloudsmith Brings AI Agent Skills Under Artifact Registry Controls at KubeCon
- Cognition Turns Devin Loose on Its Own Code with GPT‑6 Astra